VARIABLES METEOROLÓGICAS QUE INFLUYEN EN EL BALANCE

We utilize two years (March 2005-March 2007) of loyalty card data for 798 customers from a leading national grocery store chain for. Fifty-five million customers are enrolled in the retailer's loyalty card program and 13M of these household’s shop at the retailer every month. Furthermore, 80% of the retailer's customers carry the loyalty card and 95% those households’

22

The information gathered on cardholders was used by the retailer to send customized and mass direct mail promotional incentives to certain households based on their value to the firm. Despite the growth of targeted digital promotions, higher response rates of direct mail (3.7%) vs. digital (0.6%) promotions explain why retailers continue to rely on direct mail promotions to retain customers and increase store spending (Forbes Insights and Turn 2015). The dataset contains information about the timing and nature of the direct mail efforts, however details about how households were chosen for each promotion was not disclosed by the data provider, other than those chosen to receive each promotion type crossed a certain threshold of revenue contribution. We note that we test for the potential of sample selection bias later in the paper.

The customized promotions contained 14-16 coupons specially chosen for each household receiving one. The coupons were of the form, “$5 off Category X when spending $20” and “$1 off Brand X”, and the mailer envelope included the copy, “Thanks for being such a great customer!

Please enjoy the savings reserved only for premier customers like you.” Figure 4 shows an example of a customized promotion.

The mass promotions took the form of a brochure with a themed event (e.g., "Back to School," Healthy Living," "Breakfast," etc.) and contained recipes, helpful tips, etc. along with approximately ten coupons for products related to the theme (i.e., pencils and notebook, health products, breakfast sausage, etc.). The same set of coupons was included in every brochure for the same themed event and the mailer read, "Premier customer coupons inside” or “Thanks for shopping with us.” Figure 4 shows a page from a brochure under the “Back to School” theme. Both customized and mass promotions have coupon redemption duration, and, in this study, we are interested in the household’s spending during this duration.

23

The dataset contained the store visits of every household (when it used its loyalty card), beginning with the first store visit after signing up for the retailer’s loyalty program. Therefore, the first direct promotion observed for each household was the first direct communication between the retailer and the household. We observe the total amount a household spent at the retailer on each store visit, the primary variable of interest and various other campaign-related variables, loyalty program variables and customer demographic information. Our dataset has a total of 798 unique customers whose purchase history was recorded by the retailer within a two- year period using the loyalty card.

Figure 3 Customized (left) and Mass Coupon (right)

Since some customers were enrolled in the loyalty program towards the later part of the observation period, there is a difference in the time frame for which the transaction data of the customer is recorded. On average, we have 92 weeks of total store expenditure and visits history for these customers with the maximum being 102 weeks and minimum being 25 weeks. The complete list of variables and their description can be found in the Table 1 below.

24

Table 1 Variable Description

Variable Description

Dependent Variable

Campaign Sales (SALES) Total amount of dollars spent during the promotional campaign. Includes a reference six-week period when no campaign was present.

Campaign Variables

Campaign (CAMP) Represented by two dummy indicators such that CAMP1=

1 (𝑖𝑓 𝑡𝑎𝑟𝑔𝑒𝑡𝑒𝑑 𝑐𝑎𝑚𝑝𝑎𝑖𝑔𝑛 𝑖𝑠 𝐴) and CAMP2 = 1 (if

targeted campaign is B) during the observation period. ‘No Campaign’ being the baseline level.

Campaign Before (CBEF) Number of times the household is targeted with the similar campaign before.

Campaign Period (CPER) Number of weeks for which the household is exposed to a promotional campaign

Total Coupons (COUP) Number of coupons sent to the household if targeted with a promotional campaign

Favorite Coupons (FAVC) Percentage of coupons in the household’s favorite category Total Campaigns (TCAMP) Total number of campaigns the household is exposed to

during the promotion period.

Loyalty Program Variables

Loyalty discounts (LDISC) Average loyalty discounts advertised in the household’s favorite category during the promotion period.

Weekly Mailer (MAIL) Average number of household’s top 30 favorite categories featured in the weekly mailer during the promotion period End-of-Aisle Display (DISP) Average number of household’s top 30 favorite categories with an end-of-aisle display during the promotion period. Time in Loyalty Program (TIME) Number of weeks household is enrolled to the loyalty

program when received the campaign

Customer Demographics

Age (AGE) Age of the customer

Income (INCOME) The income in dollars of the customer

Married (MARRIED) Dummy variable that equals 1 if a customer is married and 0 otherwise

25

Table 2 Variable Description continued

Household Size (HHSIZE) Total number of family members in the household

Transaction Variables

Recency (REC) Weeks elapsed since last purchase Frequency (FREQ) Number of store visits in past 20 weeks

Monetary (MON) Monetary value of spending during past 20 weeks pcs.2 Second principle component score of the spending path vcs.2 Second principle component score of visits path

Over this two-year period, the retailer targeted the customers with 24 direct promotions with some customers receiving up to 14 different direct promotion while some are receiving none. These promotional campaigns spanned from 6 to 10 weeks implying that the customers had the opportunity to redeem the coupons received in the direct mail within that duration. There were durations within the observation period when customers had the opportunity to redeem coupons from both types (customized and mass) of direct promotions or were targeted with multiple promotional campaigns during the same time.

Out of 24 direct promotions, 3 were customized while the remaining 21 were mass promotions where all customers received the same set of coupons based on a particular theme. However, customized promotions usually in our sample, the least amount of time it took for a customer to receive a direct promotion discount since the beginning of the loyalty program was 21 weeks. Therefore, for each customer, we utilize 20 weeks of prior spending and visits trajectory to understand the response to direct promotions. Households tend to exhibit weekly and monthly cyclic fluctuations in store expenditures (Damon, King and Leibtag 2013), therefore a twenty- week spending history is sufficient to capture key patterns in the spending paths. Furthermore, due

26

to the frequent targeting of customers by the retailer in our data, it was rare for a household to go more than ten weeks without being targeted with a promotion.

Table 3 Variable Summary

Campaign A N = 2172 Campaign B N = 1667 No Campaign (NC) N = 708

Mean SD Mean SD Mean SD

SALES 553 399 562 369 479 400 CBEF 3.26 2.06 4.94 3.15 0.70 1.68 FAVC 19.19 11.11 6.05 11.04 0.00 0.00 ALLC 1.58 0.70 2.29 0.78 0.00 0.00 COUP 7.56 2.38 9.30 6.83 0.00 0.00 LDISC -19.50 3.32 -19.62 3.34 -19.28 3.50 DISP 18.50 3.71 18.94 3.65 18.06 4.56 MAIL 20.01 3.04 20.45 3.02 19.92 4.10 AGE 44.49 11.68 42.83 11.12 44.89 11.63 INCOME 64.50 47.49 71.12 51.95 64.85 47.73 MARRIED 0.42 0.49 0.45 0.50 0.44 0.50 NUMKIDS 0.53 0.94 0.67 1.01 0.57 0.97 TIME 58.59 14.76 65.71 16.92 39.20 19.31 CPER 8.22 0.57 6.43 0.99 8.05 1.38 REC 0.41 1.18 0.27 0.82 0.77 2.36 FREQ 15.67 3.54 16.72 3.09 13.72 4.88 MON 67.18 42.77 89.68 50.15 54.63 44.70 pcs.2 -1.84 62.25 -2.95 72.62 12.63 63.71 vcs.2 -0.174 0.425 -0.262 0.382 0.115 0.50

The unit of analysis in our study is a customer-campaign combination and the objective is to study that impact of pre-promotion spending and visits trajectory and various loyalty program, campaign and household related factors on the spending during the promotional period. To contrast the promotional spending with non-promotional spending or to develop a baseline spending case we randomly selected a duration (uniformly drawn [6-10]) within the customer time series when no promotion was present and extract the 20-week prior spending and visits trajectory for model building. Eventually, we obtain 4547 observations for each household-campaign combination

27

(including no campaign period) and the table below summarizes the variable statistics for the dataset.

For each promotional or no promotional period we implement the proposed two-stage FDA model to extract the trajectory information in the form of the Principal Component Scores. Essentially, for all customer-campaign combination we proceed through steps of FDA framework described before to incorporate entire information in the final model. To address the research questions of the study we model the total spending of a household during a promotional campaign as following: LOG(𝑆𝐴𝐿𝐸𝑆_𝑖𝑡) = 𝛽₀ + 𝛽₁ x 𝐶𝐴𝑀𝑃_𝑖𝑡 + 𝛽₂ x 𝐶𝐵𝐸𝐹_𝑖𝑡 + 𝛽₃ x 𝐶𝑃𝐸𝑅_𝑖𝑡 + 𝛽₄ x 𝐶𝑂𝑈𝑃_𝑖𝑡 + 𝛽₅ x 𝐹𝐴𝑉𝐶_𝑖𝑡 + 𝛽₆ x 𝑇𝐶𝐴𝑀𝑃_𝑖𝑡 + 𝛽₇ x 𝐿𝐷𝐼𝑆_𝑖𝑡 + 𝛽₈ x 𝐿𝑀𝐴𝐼𝐿_𝑖𝑡 + 𝛽₉ x 𝐷𝐼𝑆𝑃_𝑖𝑡 + 𝛽₁₀ x 𝑇𝐼𝑀𝐸_𝑖𝑡 + 𝛽₁₁ x 𝐴𝐺𝐸_𝑖𝑡 + 𝛽₁₂ x 𝐼𝑁𝐶𝑂𝑀𝐸_𝑖𝑡 + 𝛽₁₃ x 𝑀𝐴𝑅𝑅𝐼𝐸𝐷_𝑖𝑡 + 𝛽₁₄ x 𝐻𝐻𝑆𝐼𝑍𝐸_𝑖𝑡 + 𝛽₁₅ x 𝑅𝐸𝐶_𝑖𝑡 + 𝛽16 x 𝐹𝑅𝐸𝑄𝑖𝑡 + 𝛽17 x 𝑀𝑂𝑁𝑖𝑡 + 𝛽𝑆 x 𝑆𝑝𝑒𝑛𝑑𝑖𝑛𝑔 𝑇𝑟𝑎𝑗𝑒𝑐𝑡𝑜𝑟𝑦𝑖𝑡 + 𝛽𝑉 x 𝑉𝑖𝑠𝑖𝑡𝑠 𝑇𝑟𝑎𝑗𝑒𝑐𝑡𝑜𝑟𝑦𝑖𝑡 + 𝜀𝑖𝑗 ; i = 1,…n, j = 1,…. 𝑚𝑖 --- 4

Where, 𝑆𝐴𝐿𝐸𝑆_𝑖𝑡is the sales or total spending for customer ‘i’ in the promotional campaign at time ‘t’, n represents the total number of households, 𝑚_𝑖 is number of repeated observations or promotional campaigns for the customer ‘i’ in the sample. It should be noted that campaign

assignment to the customers is not random and is driven by a process which is not known to us. Hence, we account for the endogeneity in the promotion assignment using propensity score weighting procedure.